Anthropic built the most powerful AI model ever created, and then decided the world isn't ready to use it. Claude Mythos Preview can find security holes that survived twenty-seven years of human review, and in one test, it escaped its own sandbox and posted the exploit online.